Methods to Expand Cell Signaling Models Using Automated Reading 149Fig. 3.Relationship between reading output and model. (a) The literature reading-
assembly flow with example. (b) Example baseline model (solid arrows) and new inter-
actions extracted by automated reading (dashed arrows). The circled numbers represent
classification described in Sect.2.2(Color figure online)
3.1 Interaction Map Extension
Each interaction of extension type can be regarded as a candidate new edge
in the model’s interaction map. LetEextbe the set of interactions provided
by reading. Suppose the baseline model isG =(V, E). Each new candidate
model can be obtained by adding a group of selected edgesEnew∈Eextand its
corresponding elements, that is,G′=(V′,E′), whereE′=E∪Enew. However, it
is impossible to enumerate all configurations of whether or not to add a new edge,
as the number of candidate models will become extremely large. For example,
if there are 100 new interactions extracted by the automated reader, there are
2100 possible extensions of the model. This number is impossible to handle,
therefore, we need heuristic methods to search for suitable configurations of
model extensions.
A possible way to tackle the issue of large number of model extension config-
urations, is to list the elements of interest in the baseline model, and include, as
an extension, interactions that are related to those elements. The set of model
elements of interest can be defined by user, depending on the questions asked or
hypotheses tested. Still, the extension configurations need to be constructed in
a systematic manner. Here we introduce the concept of ‘layer’, where layerS 0 is
the set of elements of interest. The next layer,S 1 , is the set of direct parents of
elements inS 0 , and in general,Siis the set of direct parents of elements inSi− 1.
Elements inS 1 are direct regulators ofS 0 , and thus, the extensions including
elements inS 1 may influence the model. Using this concept, we propose four
different methods to create extension configurations.
Cumulative parent set with direct extensions (CD):In this method, we
define the number of layer,n, that we want to consider, and include all new
interactions that affect any element from layer 0 up to layern. In other words,
we add an extensione=(u, v) to the model when at least one of the nodesu
andvis mentioned in layersS 0 toSn. Figure 4 b demonstrates the result of this
method wherenequals 1. Starting fromS 0 ={A, B, E}, we find its direct parents
S 1 ={C, D}. The edges in the figure represent the union of layersS 0 andS 1.